TrueBlue Talks: Machine Learning

First introductory talk about machine learning in TrueBlue.

Pre-requisites:

• Python basic knowledge
• Conda installed
• Would be nice if everyone has the environment already installed, it takes time

Please watch the intro at ML-Tour

AI - ML - DL How they are related

Machine Learning vs Deep Learning on features

We will see the concept of features later on. For this moment consider it as the R/G/B channels values in an image, for each pixel in the early layers, edges or lines in the following, corners or line crossing in the middle layers, objects near the final layers and real world targets in the very final layer.

Machine Learning Problems

There are a variety of problems that can be tackled with ML, each have a different approach and resolution technique.

What we will see today

Two examples:

• Unsupervised
• Supervised: simple Neural network VS convNet

We will have to skip Recurrent NN, time series and general AI problems. Maybe in one of the next episodes. Q-Learning, Capsule Network and Reinforcement Learning will be a nice topic, unfortunately I don’t have time to prepare them in the foreseable future as they are far more complex to explain and setup.

Unsupervised exmaple: kNN

It’s one of the most easy to understand algorithm, and gives satisfaction! we will classify access to servers and see where anomalies lies.

Supervised exmaple: images

What are we trying to solve

Our dataset

MNINT is a dataset with 10 integers to classify, too easy for our pourposes. MNIST is the classical example here, but is has became too easy to solve and reach 99.7% accuracy. I’ve found MINST-fashion where we will classify images of clothes, shoes etc. The performances here drops a bit so we will have more fun trying to solve it.

Mandatory Debunking

Myth 1

Facebook chatbots created an unkown language, AI will kill us all!

Truth

Cool description of a bug

Myth 2

AI will replace humans work

Truth

Same relation as PC VS human calculators, tractors VS horse powered agricolture VS manual agricolture etc… It’s just another tool to make us more efficient and find insights otherwise hard to discover.

Myth 3

Simpler models are more accurate

Truth

This is a misintepretation of the Occam Razor principle, where basically, to solve a statistical problem, you should make as few assumption as possible. This is true, until it grant the same level of performances. The truth is that with massive amount of data is hard to know exactly which assumption will give real help to our model or not. The root cause here is that very complex models are hard to comprehend and to modify.

Myth 4

Everyone can do Machine Learning

Truth

It’s like saying that anyone can do doctor, engineer or politician. It takes years of practice, experience and practice..I classify myself definitely as a newbie after 3 years of study and projects.

Myth 5

You always need a google-esque amount of data

Truth

Sometimes that’s true, but for most applications you can exploit Transfer learning, where you use a pre-trained massive network and then replace only the last layers with your business needs/classes/etc…

…honestly this list will be endless